🤖 AI Summary
This paper addresses extinction-level risks—such as objective misalignment, geopolitical conflict, and malicious misuse—posed by premature development of Artificial Superintelligence (ASI). Methodologically, it proposes a U.S.–China–centered international governance framework featuring a novel verifiable technical architecture that integrates FLOP-based training thresholds, end-to-end AI chip lifecycle tracking, and model usage behavior attestation—embedded within legally binding prohibitions and multilateral negotiation mechanisms to balance security and feasibility. Its primary contribution is the first ASI development constraint pathway that is technically enforceable, independently verifiable, and politically acceptable amid profound trust deficits—effectively halting high-risk capability jumps while preserving beneficial AI applications. The framework offers an operationally viable policy paradigm for global AI safety governance; however, its implementation hinges on sustained great-power political consensus and adaptive alignment with technological evolution.
📝 Abstract
Many experts argue that artificial superintelligence (ASI), if developed prematurely, poses catastrophic risks including human extinction from misaligned ASI, geopolitical instability, and misuse by malicious actors. This report proposes an international agreement to prevent the premature development of ASI until AI development can proceed without these risks. The agreement halts dangerous AI capabilities advancement while preserving access to current, beneficial AI applications. The proposed framework centers on a coalition led by the United States and China that would restrict the scale of AI training and dangerous AI research. Due to the lack of trust between relevant parties, verification is a key part of the agreement. Limits on the scale of AI training are operationalized by FLOP thresholds and verified through the tracking of AI chips and verification of chip use. Dangerous AI research--that which advances toward artificial superintelligence or endangers the agreement's verifiability--is stopped via legal prohibitions and multifaceted verification. We believe the proposal would be technically sufficient if implemented today, but advancements in AI capabilities or development methods would hurt its efficacy. Simultaneously, there is not yet political will for such an agreement. Despite these concerns, we hope this agreement can provide direction for AI governance research and policy.